Daily Twitter Digest

Mathematics

Study: Lean Verification of OpenAI's Navier-Stokes Proof Doesn't Match the Written Argument

The paper argues that AI-driven autoformalisation—translating natural-language math proofs into formal languages like Lean for mechanical verification—cannot in general be trusted to faithfully capture the original argument. The authors claim the task of resolving ambiguity in mathematical natural language is arbitrarily hard within the arithmetical hierarchy (harder than the Halting problem), and illustrate this with concrete examples, including OpenAI's announced Lean-verified proof of blow-up for Navier-Stokes solutions, where they show the formalized Lean statement does not actually correspond to the natural-language proof. The broader claim is that a Lean verification passing is not sufficient evidence that the informal proof it's supposedly checking is correct. Commentary on Twitter focused heavily on the Navier-Stokes case as a concrete, high-profile embarrassment for AI-assisted formal verification claims, with some noting the irony that another AI model reportedly fixed the mismatch. Other users raised skeptical/cynical asides about how such subtle mistranslations could go undetected in large formalized codebases, questioning the practical trustworthiness of AI autoformalisation pipelines at scale.

Discussion: 6 tweets from 6 authors · @Quasilocal, @pmddomingos, @mike_mates, @semiDL, @LogicPractice +1

Social Science

Study Maps Far-Right's Educational Network in Spain via Social Media Analysis

This study investigates the political-pedagogical ecosystem linked to neofascism in Spain, focusing on its relational infrastructure in education rather than just its electoral rise. Using Social Network Analysis (UCINET, Gephi) and digital ethnography on 150,000 posts related to the #StopLeyCelaá hashtag (opposing the LOMLOE education law), the researchers mapped the actors involved. They identify a complex network of parties, foundations, religious groups, and media that promote market logic in education, parental primacy, and a national-patriotic curriculum while opposing diversity content. On Twitter, researchers shared the findings as the first study to map this specific "educational map" of the far-right in Spain, highlighting how diverse political and religious factions are interconnected through specific individuals and platforms to spread these narratives.

Discussion: 2 tweets from 2 authors · @nolo14, @EnriqueJDiez

Computer Science

PivotOPD Trains Agents to Recover from Early 'Pivotal' Mistakes

The paper studies on-policy distillation for multi-turn language agents, finding that over half of failed rollouts stem from a single early 'pivotal mistake' that compounds across turns but is often still recoverable with brief guidance. PivotOPD addresses this by combining reverse-KL distillation to steer students away from pivotal errors with forward-KL distillation on teacher-provided recovery actions for the turns immediately after a mistake occurs. Tested against 13 baselines on ALFWorld, WebShop, and search-based QA, it improves Qwen3-1.7B and 8B students, and also boosts a Nemotron-3.5 model's SWE-Bench Verified resolve rate by 3.2%. Twitter discussion highlighted the headline statistic that recovery rates jumped from 8.3% to 72.7%, framing the method as teaching agents 'the way back' from compounding errors without needing a larger teacher model.

Discussion: 2 tweets from 2 authors · @ahatamiz1, @yinghui_he_

Biology

New Automated Training Paradigm Finds Long-Term Memory in Planaria Survives Head Regeneration

The paper reports a computerized training and testing system for planarian flatworms designed to avoid confounds that plagued earlier manual conditioning studies. Using an environmental familiarization protocol, the authors find that worms retain memory for at least 14 days—long enough for a severed head to regrow—and that decapitated, trained planaria show faster relearning (a 'savings' effect) after regenerating a new head, suggesting some memory trace persists through brain regeneration. The authors propose planaria as a tractable model for studying how memories are encoded and retained across radical tissue remodeling, with potential relevance to regenerative neuroscience. Twitter discussion contrasted this with the long and troubled history of planarian memory research, noting that McConnell's famous 1960s light-shock conditioning experiments were never reliably replicated, and that a more recent attempt by Gershman's lab reportedly failed to get worms to learn that task at all—framing this new automated approach as the latest attempt to put the field's claims on firmer empirical footing.

Discussion: 1 tweets from 1 authors · @enjoyingthewind

Computer Science

Base Models' Reasoning Can Be Unlocked by Simple Starting Token Cues

The paper shows that base language models' reasoning performance is strongly tied to the specific tokens that begin their response—fixing a cue like ".\n\nOkay" boosts Olmo-3-7B's MATH-500 accuracy from 42% to 78%, nearly matching RL-trained versions, and "Alright," similarly lifts Qwen3-14B. The authors argue RL mainly works by making models more likely to produce these effective cues themselves, and show via causal data interventions that cue effectiveness can be transplanted onto arbitrary words (even "chicken") by editing training data, tracing the effect back to associations with specific document types—including effects on refusal/compliance safety behavior. The tweet accompanying the paper points to an interactive blog post letting readers explore how token cues and RL affect base model reasoning, with code released alongside; the post itself didn't generate deeper critical discussion in the thread shown.

Discussion: 1 tweets from 1 authors · @SophieLWang

Social Science

Book Argues LDP Keeps Power by Rewarding Loyal Voter Blocs with Budget Allocations

Amy Catalinac's book contends that Japan's dominant Liberal Democratic Party sustains its decades-long electoral dominance not through popularity but through strategic resource allocation: in districts with identifiable voter groups, LDP politicians channel central government spending to reward blocs that vote for them, forcing groups to compete for resources. Drawing on quantitative and qualitative data from 1980-2014, the argument also explains persistent opposition weakness and why Japan's 1994 electoral reform failed to disrupt this dynamic. A Japanese political scientist tweeted that the book's intra-district findings—that municipalities with higher LDP vote shares receive larger budget allocations—resonated directly with a specific contemporary case (Tochigi 3rd district), framing it as empirical confirmation of the clientelist mechanism Catalinac describes, contrasted with their own inter-district analysis.

Discussion: 1 tweets from 1 authors · @000fukumoto

Medicine

Meta-Analysis Links Cannabis Use to Doubled Risk of Manic Symptoms, Bipolar Disorder

This systematic review and meta-analysis (published in Journal of Affective Disorders) pools existing studies to examine whether cannabis use predicts the later onset of manic symptoms or bipolar disorder; no abstract is available, so the paper's precise methodology and effect sizes beyond the headline finding are unconfirmed here. According to a Twitter summary of the paper, cannabis use was associated with more than double the odds of subsequently developing bipolar disorder or experiencing a manic episode. The tweet (in Spanish) presented this as a notable warning about cannabis-related psychiatric risk, though no substantive methodological critique or skepticism was raised in the discussion, which was limited to a single highly-engaged post sharing the finding.

Discussion: 1 tweets from 1 authors · @metodogolo

Mathematics

Arxiv Paper Claims Full Classification of Primes in Lubin-Tate Space, Twitter Cries Foul

The paper claims to prove, for every height n and prime p, that stable prime ideals of the ring A_n = W(F_{p^n})[[u_1,...,u_{n-1}]]/(p) under the Morava stabilizer group action are exactly the height ideals, extending this to a related ring R_n. The authors argue this yields the Hovey-Strickland classification of thick tensor ideals in K(n)-local spectra, building the proof via a mix of Cartier theory, Fargues-Fontaine bundles, and Chai's rigidity theorem. On Twitter, a widely shared comment called the paper an instance of "intellectual fraud" on an "incomprehensible" scale, though the discussion gave no specific technical rebuttal, reflecting broader suspicion (voiced elsewhere about similar arxiv submissions) that such densely jargon-laden proofs may be AI-generated or unverified rather than rigorous mathematics.

Discussion: 1 tweets from 1 authors · @doomslide

Chemistry

Japan Academy Prize Article Recaps Soai's Asymmetric Autocatalysis Reaction

No abstract is available for this short Japanese-language article, so this summary is based on discussion only. The piece, written by Kawasaki, reportedly recaps the achievements of Kenso Soai, discoverer of the "Soai reaction" — an asymmetric autocatalytic reaction in which a chiral product catalyzes its own formation, amplifying small initial enantiomeric imbalances into high optical purity, work recognized with the Japan Academy Prize. The tweet highlights that the article is written in a compact, accessible style, and notes the striking claim that the Soai reaction was singled out among papers published in Nature over the 50-year span from 1950 to 2000, underscoring its significance in origin-of-chirality research.

Discussion: 1 tweets from 1 authors · @yuuri_eguchi

Mathematics

Empirical Bayes g-modeling meets Mixture-of-Experts in new estimation framework

The paper extends empirical Bayes g-modeling—where a latent prior over regression coefficients is estimated nonparametrically from data—to let that entire prior depend on covariates, rather than being shared across all units. The authors show the unrestricted version of this problem is ill-posed, and resolve it by introducing an 'infinite Mixture-of-Expertizing' class of conditional priors, proving that the resulting nonparametric MLE is exactly a softmax-gated Mixture of Experts whose number of experts emerges from the data rather than being fixed in advance; they back this with Lipschitz continuity and gate-approximation theory, plus tests on synthetic and real datasets. The Twitter discussion, from one of the authors, simply highlights the paper's core novelty: connecting 1950s-era empirical Bayes theory to the Mixture-of-Experts architecture now familiar from large language models, with no substantive critique offered.

Discussion: 1 tweets from 1 authors · @kyanostat

Medicine

Network Meta-Analysis Compares Antifungal Prophylaxis Options in Blood Cancer Patients

This systematic review and network meta-analysis evaluates antifungal prophylaxis strategies for patients with haematological malignancies undergoing chemotherapy or stem cell transplantation, aiming to compare the relative effectiveness of different agents in preventing invasive fungal infections. No abstract was available, so details on specific drugs compared, outcome measures, and conclusions are not confirmed here; this summary is based on the title and limited discussion. The Twitter commentary consists of a single share of the paper without added analysis or critique, so no substantive discussion points or skepticism have emerged yet.

Discussion: 1 tweets from 1 authors · @AliSMV7

Computer Science

NEEDLE Stitches Robot Demos into Better Training Data Without New Interaction

The paper introduces NEEDLE, an offline algorithm that improves robot imitation-learning datasets by inserting short, verified 'action bridges' between recorded observations—using only RGB images, proprioception, and episode-level success/failure labels, with no new environment interaction or privileged state. This lets suboptimal detours be bypassed and failed trajectories repurposed, while a new sampling scheme incorporates these stitched connections into policy training without discarding original demonstrations or synthesizing intermediate images. The authors report an average 21 percentage-point improvement in success rate over the strongest baseline across real-robot tasks. The sole tweet introducing the work frames it as addressing why trajectory stitching works for high-dimensional robot data, not just whether it can be made to work, but offers no independent critique or discussion yet.

Discussion: 1 tweets from 1 authors · @JuntaoRen

Biology

Lineage Tracing Shows Mouse Embryo Cells Lose Potential Asynchronously

The preprint combines DNA Typewriter lineage recording with tetraploid complementation to reconstruct single-cell genealogies in mouse embryos and stem-cell-derived embryo models, introducing an analysis framework called FateVec to infer fate-bias dynamics from these phylogenies. The authors report that loss of developmental potential is progressive and asynchronous—some lineages show fate bias well before gastrulation, neural crest cells remain multipotent longer than expected with variable restriction timing across founder clades, and cells sharing transcriptional identity can have distinct lineage histories, including unexpected embryonic contributions to extraembryonic-like states. The central claim is that cell identity alone cannot predict a cell's developmental ancestry or when its fate becomes restricted. On Twitter, lead author Magdalena Zernicka-Goetz highlighted the finding that identical cell states can mask divergent developmental histories, framing the work as linking lineage, fate timing, and gene expression; discussion so far is limited to this framing with no substantive criticism yet surfaced.

Discussion: 1 tweets from 1 authors · @ZernickaGoetz

Chemistry

CrystalJev Turns Atomistic Foundation Models into Fast, Calibrated Decision-Makers

The paper argues that atomistic foundation models, typically used as slow brute-force simulators whose thresholded energies are taken at face value, are better understood as fast decision-makers. CrystalJev queries a frozen interatomic potential with a single forward pass per unrelaxed structure and converts this into calibrated probabilities with finite-sample guarantees, plus a value-of-information rule deciding when fuller relaxation is actually needed. Tested across 65 Matbench Discovery models and validated against 700 new DFT calculations, the method reportedly matches relaxation-based decisions at about a thirtieth of the computational cost, while flagging that single-pass forecasts can modestly overstate the stable fraction of novel candidates (by up to 2.1 percentage points in the prospective test). The one tweet discussing the paper (in Japanese) highlights the speedup as impressive, framing the approach as applying calibrated, fast 'System 1' decision-making to materials discovery and noting the ~30x reduction in compute cost — with no substantive criticism offered.

Discussion: 1 tweets from 1 authors · @yoko_materialDX

Social Science

Legal Scholar Argues Misgendering Should Count as Workplace Harassment in SA Law

This article contends that persistent or deliberate misgendering at work should be legally recognised in South African labour law as harassment, and in some cases misconduct warranting dismissal. The argument draws on the Constitution, the Employment Equity Act, the Harassment Code, and the Labour Relations Act, while distinguishing good-faith mistakes from intentional or reckless conduct continued despite known harm. It closes with proposed reforms to workplace policy, training, disciplinary procedure, and dispute resolution. The Twitter discussion consists mainly of the author announcing publication, with no substantive critical commentary visible in the available tweets.

Discussion: 1 tweets from 1 authors · @mx_mokgoroane

Medicine

Erratum to German Medical Chapter on Chronic Fatigue Syndrome Sparks Controversy

This is a publisher's erratum to a book chapter on Chronic Fatigue Syndrome (ME/CFS) published by Springer Nature; no abstract is available, so this summary is based on Twitter discussion only. The erratum apparently revises or corrects value-laden statements about the condition, though the exact content and scope of the correction remain unclear from available sources. Commentators note the erratum claims certain judgmental or contentious statements were only introduced during the production process after review, rather than being present in the original peer-reviewed text. On Twitter, commentary (notably from @GrandeBettina, tagging #MECFS and #FBM26) expressed skepticism and criticism, calling the publisher's explanation 'absurd' and questioning why a reputable academic publisher like Springer Nature would be associated with what they characterized as a dubious episode in science publishing. The criticism centers on distrust of the stated timeline for how disputed content entered the chapter, rather than on the scientific content of ME/CFS research itself.

Discussion: 1 tweets from 1 authors · @GrandeBettina

Other

Amateur Analysis Flags TIC 4206066 as Possible Host to Two Planets

A Zenodo-hosted write-up explores a 3D model of the star TIC 4206066 built from TESS photometry, proposing that two transit-like signals in the data could indicate the presence of two planets. The author frames this as a tentative hypothesis rather than a confirmed detection, releasing the underlying data, code, and an interactive 3D visualization for independent scrutiny. No peer-reviewed abstract beyond this descriptive summary was provided, so claims about planetary nature remain preliminary. On Twitter, the main poster was notably cautious themselves, stating that "none of this confirms it yet" and proposing a concrete test: whether the star dims on schedule in November would make the case "a lot more serious," while a non-dimming result would also be reported. This pre-registered, falsifiable framing drew attention as an example of transparent amateur exoplanet analysis rather than a finished scientific claim.

Discussion: 1 tweets from 1 authors · @p_rabtsevich

Medicine

Study: 10g/hr Carb Intake Fails to Prevent Blood Glucose Drop During Exercise

Researchers had eight trained male cyclists ride for 3 hours at four different carbohydrate intake rates (10, 90, 120, 150 g/hr) followed by a high-intensity capacity test, using labeled glucose-fructose drinks to track substrate oxidation. They found that 10 g/hr resulted in lower blood glucose and shorter exercise duration than all higher doses, while 150 g/hr offered no additional benefit over 120 g/hr for whole-body carbohydrate oxidation; the authors estimate an optimal dose near 107 g/hr but note substantial individual variation, supporting personalized carb intake recommendations. On Twitter, one of the researchers highlighted the finding as clear evidence against the claim that 10 g/hr is sufficient to prevent exercise-induced hypoglycemia, framing the result as a straightforward empirical refutation of that absolute claim.

Discussion: 1 tweets from 1 authors · @Gonzalez_JT

Chemistry

New Magnesium Zincate Compound Releases Hydrogen-Methane Blend at 80°C

Researchers report Mg[(CH3)ZnH3], a crystalline, hydride-rich organometallic compound free of bulky supporting ligands, that serves as a single-source solid-state precursor for hydrogen-methane fuel blends. Under solvent-free, anaerobic conditions, the compound undergoes reductive elimination at temperatures as low as 80°C, releasing both gases and leaving behind elemental magnesium and zinc. The authors frame this as a proof-of-concept for molecular precursors capable of delivering mixed fuel gases from a single compound, relevant to solid-state storage of transitional fuel blends. The Twitter discussion is minimal, with one post highlighting the compound as a notable fundamental discovery with potential practical energy applications, but offering no substantive technical critique.

Discussion: 1 tweets from 1 authors · @MainTvm

Computer Science

QuadTok Uses Quadtree Structure to Cut Image Tokens by ~10%

QuadTok proposes a hierarchical quadtree tokenizer for image generation that adaptively allocates more tokens to visually complex regions and fewer to homogeneous ones, bridging 2D grid and 1D sequence tokenization approaches. The authors report their ImageNet-trained tokenizer saves roughly 10% of tokens versus a fixed 256-token grid while preserving reconstruction fidelity, and a 947M-parameter GPT-style autoregressive model built on this structure achieves 2.08 gFID on ImageNet 256×256, with the tree topology also enabling zero-shot spatially controlled generation. The lead author frames hierarchical representations as a promising general direction for vision, with this work as an initial step toward that goal. The Twitter discussion so far consists mainly of the author's own thread introducing the paper and code, emphasizing the hierarchical/adaptive tokenization angle as the key contribution; no independent critical commentary or skepticism has yet surfaced in the available tweets.

Discussion: 1 tweets from 1 authors · @MatthewMao